I built the same game on Roblox, Unity, and the web. Only one didn't fall apart.

Andrej Karpathy’s ‘vibe coding’ concept—using AI as a high-level programming language where you barely touch the code—sounds like magic for game development. It kind of is, right up until the project hits a few hundred lines. Then the context window fills up, model performance degrades, and your game turns into digital spaghetti.

That’s the core problem VibeGame’s creator ran into while testing AI-driven game development across three platforms. The experiment was straightforward: build the same incremental game, inspired by Grass Cutting Incremental, using Claude Code on Roblox, Unity, and a bare-bones web stack. The results weren’t close.

Roblox’s official MCP server offered beautiful high-level abstractions—AI instantly grasped the syntax—but the walled garden killed it. No file system access meant no real context management. Unity’s unofficial MCP was the opposite problem: full file access, but the engine’s endless ways to accomplish the same task, which shift constantly across versions, left the AI confused and unreliable. “There are many ways to do everything in Unity,” the developer noted, which means you spend more time specifying how to do things than what to do.

The web stack—three.js for rendering, rapier for physics, bitecs for game logic—blew both away in AI performance. The reason is obvious: massive training data from web development. But there’s a catch. These libraries are so low-level that you end up building an engine before you can build a game. The fix was creating a custom high-level abstraction layer on top, essentially an AI-native game framework that keeps projects lean enough to fit in a model’s context window.

To manage what context remains, the developer built Shallot, a lightweight system for Claude Code with two commands: /peel to load context at the start of a conversation, and /nourish to update it at the end. It’s dead simple and intentionally unopinionated. The bigger lesson here isn’t about any single tool. It’s that vibe coding games demands a platform that enforces discipline—high-level abstractions the AI understands well, minimal boilerplate, and a project structure that doesn’t metastasize. Right now, that platform doesn’t quite exist off the shelf. But the blueprint is getting clearer.